Supplementary Materials

Exploring High-Dimensional Spaces: Evaluation of a Node-Based Interface for Neural Audio Synthesis

Leonardo Auri¹ · Tom Beckmann² · Henrik von Coler³ · Robert Hirschfeld²
¹Audio Communication Group, Technische Universität Berlin · ²Hasso Plattner Institute, University of Potsdam · ³School of Music, Georgia Institute of Technology

This page collects the supplementary material accompanying the paper: a live build of the interface running in the browser; the complete source of the interface, backend and analysis; the synthetic dataset one of the neural decoders was trained on, together with the code that generates it; and all raw data collected from the user study.

Live demo #

The proposed node-based interface and the slider baseline, running in the browser with the neural decoder executing locally — no audio or usage data leaves your machine.

fulldim-interface in the browser web build · 166 MB

Choose New Node Scene or New Slider Scene to try either condition directly.

Needs a current desktop browser with WebGL 2 and WebAssembly; the decoders are downloaded on first load, so give it a moment. Audio starts after the first click, as browsers require.
Open the demo

Interface, backend and analysis code #

The node-based interface and the slider baseline, the inference backend driving the RAVE decoder in real time, and the analysis scripts that produce every statistic and figure reported in the paper.

fulldim-interface source · GitHub

The publication branch is the frozen state of the repository as used for the experiment and the reported analysis.
github.com/leoauri/fulldim-interface @ publication

Summed sine dataset and generator #

Participants could choose between two RAVE decoders. One was trained on GTSinger; the other on the synthetic dataset of summed sine tones published here, which yields a timbral space of inharmonic, spectrally varied sounds.

mixtones — training dataset ZIP · 819 MB

430 WAV files, 1.71 GB unpacked. Mono, 44.1 kHz, 32-bit float. Every clip layers a random number of sine tones with independently randomised frequency, amplitude and delay–attack–sustain–release envelope, passed through a reverb tail and normalised. The set was written in two runs: 000000–000399 at 2²⁰ samples (23.8 s) and 000400–000429 at 2¹⁸ samples (5.9 s).
mixtones.zip
Two examples from the dataset WAV · 0:23 each

Converted to 16-bit for playback in the browser; the dataset itself is 32-bit float throughout.

000000.wav

000200.wav
tone-datagen source · GitHub

The generator used to produce the dataset above. Generation is unseeded, so running it reproduces the characteristics of the dataset rather than the exact files — the clips used for training are those in mixtones.zip.
github.com/leoauri/tone-datagen @ publish-EHDS

Raw experimental data #

Everything collected from the sixteen participants of the within-subject study, one directory per participant (p01–p16).

Experiment data, screenshots and questionnaire responses ZIP · 8.6 MB

Per participant, under experiments/:

condition_{first,second}_preset.json — the completed preset at the end of each condition, as full 128-dimensional vectors; condition_{first,second}_preset.png — a screenshot of the interface in its final state; condition_{first,second}_demo.csv — demonstration data, every vector change timestamped across all 128 dimensions; metadata.json and session.json — experiment metadata and timing data, including condition order, per-condition start, end and duration, selected model and consent record; csi_first.json, csi_second.json, csi_paired.json and goldmsi.json — questionnaire responses, each item recorded with its text, its id and the order in which it was presented.

The screenshots/ folder holds the same preset screenshots named by condition (pNN_node_scene.png, pNN_slider_scene.png) rather than by presentation order, as consumed by the analysis scripts.

Data is pseudonymised: participants appear only as p01–p16, and no demographic responses beyond the Gold-MSI items are included.
ehds-raw-data.zip